Re: Regression in 5.3 for some FS_USERNS_MOUNT (aka user-namespace-mountable) filesystems

[Al Viro <viro@zeniv.linux.org.uk>]

On Fri, Jul 26, 2019 at 03:47:02PM -0700, Linus Torvalds wrote:

> Of course, then later on, commit 20284ab7427f ("switch mount_capable()
> to fs_context") drops that argument entirely, and hardcodes the
> decision to look at fc->global.
> 
> But that fc->global decision wasn't there originally, and is incorrect
> since it breaks existing users.
> 
> What gets much more confusing about this is that the two different
> users then moved around. The sget_userns() case got moved to
> legacy_get_tree(), and then joined together in vfs_get_tree(), and
> then split and moved out to do_new_mount() and vfs_fsconfig_locked().
> 
> And that "joined together into vfs_get_tree()" must be wrong, because
> the two cases used two different namespace rules. The sget_userns()
> case *did* have that "global" flag check, while the sget_fc() did not.
> 
> Messy. Al?

Digging through that mess...  It's my fuckup, and we obviously need to
restore the old behaviour, but I really hope to manage that with
checks _not_ in superblock allocator ;-/